Think Like an AI Researcher Module 4 of 6
57.1% complete
Module 04 · 16 min read

Designing Your Own AI Study

When vendor benchmarks are not sufficient, you need an internal evaluation. This module covers sample size, confounders, controls, and what makes a study trustworthy enough to act on.

🧪 Study design framework ⚙️ Interactive sample size estimator 📋 Evaluation template

When you need your own evaluation

Vendor benchmarks measure vendor-chosen tasks. Your deployment context is not a vendor-chosen task. At some point in a serious AI procurement or deployment process, you need your own data. The question is how to generate it in a way that is actually trustworthy.

An internal AI evaluation does not need to meet the standards of an academic study. It does need to meet a simpler threshold: the results should be stable enough that a different team member running the same evaluation would reach the same conclusion.

The evaluation threshold

Your evaluation needs to be trustworthy enough to act on, not perfect enough to publish. The minimum viable standard: if you ran the evaluation again next week on a different sample, would the conclusion change? If yes, the evaluation is too noisy to decide anything.

Step 1: Define the task with precision

The most common internal evaluation failure is a vague task definition. "Test how well the model handles customer emails" is not a task. "Classify customer support emails into five categories: billing, technical issue, feature request, complaint, and other : with an error rate below 5% on ambiguous emails from our Q3 archive" is a task.

Before any model touches your data, write down:

01

Input specification

What exactly is the model receiving? Format, length constraints, source, any preprocessing. The more specific, the more reproducible.

02

Output specification

What is a correct output? Who decides? If human evaluators decide, how will they score? Write the rubric before seeing any outputs.

03

Success threshold

What number, produced by this evaluation, means you proceed? Define it before running the study. Post-hoc threshold selection is the evaluation equivalent of cherry-picking.

04

Failure definition

What would a fail look like? If you cannot describe failure concretely, the success threshold is not calibrated against anything real.

Step 2: Sample size

Sample size determines whether your results are stable. Too small, and the conclusion could flip with five more examples. The right number depends on the accuracy level you are trying to detect and how much variance you are willing to tolerate.

A practical rule for enterprise evaluations: if you are trying to detect a difference of 10 percentage points or more, 100-200 examples is usually sufficient. If you need to detect a 5-point difference, you need 400+. If you need to detect a 2-point difference, you are conducting academic research, not an enterprise pilot.

Sample size estimator
10 pp
75%
Common mistake

Running 20 examples and declaring a winner. At n=20, a model that is truly 10 points better will not reliably appear better due to sampling noise. You will get the wrong answer roughly 30% of the time with that sample size.

Step 3: Control for confounders

A confounder is a variable that affects your outcome measurement but is not the thing you are trying to evaluate. In AI evaluations, the most common confounders are:

ConfounderHow it appearsControl method
Prompt variation Model A gets a more specific prompt than Model B Use identical prompts for all models
Sample selection Easy examples in the test set inflate apparent accuracy Random sample from real production cases
Evaluator bias Human raters score outputs higher when they know which model produced them Blind evaluation: raters do not know which model output they are scoring
Temporal shift Test data from six months ago does not reflect current input distribution Sample from the most recent 4-8 weeks of production inputs
Temperature variation Different temperature settings change output quality inconsistently Fix all inference parameters across all model comparisons

Step 4: Measurement design

How you measure correct versus incorrect determines what your evaluation actually captures. There are three measurement approaches, each with different tradeoffs.

Automated metrics

Rule-based or model-based scoring at scale. Fast and reproducible, but can diverge from human judgment. Best for tasks with unambiguous correct answers: classification, extraction, calculation. Examples: exact match, F1 score, ROUGE.

Human evaluation

Expert raters score outputs against a rubric. Expensive and slow, but captures nuance. Required for tasks where correctness is context-dependent: summarization quality, tone appropriateness, legal risk assessment. Must use blind evaluation and inter-rater reliability checks (Cohen's kappa should exceed 0.6 for the evaluation to be credible).

LLM-as-judge

A separate model scores outputs from the model being evaluated. Faster than human evaluation, cheaper than both. Subject to positional bias (prefers the first option) and self-preference (a model scores its own outputs higher). Liang et al. (arXiv:2211.09110) document these biases in detail. If you use LLM-as-judge, use a model from a different provider than the one being evaluated, and randomize the order of options presented.

Measurement approach tradeoffs: illustrative comparison
Illustrative comparison of three evaluation approaches across four dimensions. No approach dominates; choice depends on task type and available resources. Not empirical data.

Step 5: Write the evaluation report before running it

This is the most counterintuitive step. Before you run the evaluation, write out: what success looks like, what failure looks like, and what you will do in each case. This eliminates the most common form of internal evaluation bias: deciding what the results "really mean" after you see them.

If the evaluation comes back ambiguous (results are close to the threshold), that is also informative. It means the difference between models is not large enough to justify a decision based on this evaluation. You either need a larger sample or a more discriminating task.

Pre-registration principle

Write down your decision rule before you see the data: "If Model A exceeds 80% on this task with a sample of 150 cases, we proceed to pilot. If not, we evaluate the next candidate." This removes the largest source of evaluation bias from internal studies.

Your team runs an internal evaluation comparing two document summarization models. Human raters score 60 summaries from each model. One rater scores all outputs from Model A first, then all from Model B. What confounder does this introduce?
Sample selection bias : the summaries were not randomly chosen
Temporal shift : the documents may be from different time periods
Evaluator bias : the rater knows which model produced each output and may calibrate differently after seeing 60 outputs
LLM judge bias : a second model was not used for scoring
You want to detect whether a new model is at least 8 percentage points more accurate than the current system. What minimum sample size is most appropriate?
20 examples : enough to get a directional signal
50 examples : standard for internal evaluations
150-200 examples : the minimum for reliable detection of an 8-10 point difference
500+ examples : always use the largest sample possible
Try this: design a 2-page evaluation plan
  1. Pick one AI capability your team is currently evaluating or considering (document classification, call summarization, contract review, etc.).
  2. Write the input specification, output specification, success threshold, and failure definition.
  3. Estimate the sample size you need. Use the calculator above as a starting point.
  4. List the three confounders most likely to affect your evaluation and write a control for each.
  5. Decide: automated metrics, human evaluation, or LLM-as-judge? Write the rationale.
  6. Write the decision rule before seeing any data.

Excited about AI, innovation, and growth?

Start a conversation
Was this module useful?
Before you go

You can now design an internal AI evaluation that is trustworthy enough to act on. The next module covers the organizational side: how to commission AI research from a team and what deliverables to require.

What you learned

Precise task definition, sample size estimation, confounder control, three measurement approaches, and the pre-registration principle.

Module 5 preview

Commissioning AI Research from a Team: the 10 questions that separate a credible research brief from a vague ask.